Upload config.yaml
Browse files- config.yaml +82 -0
config.yaml
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# MTP Mini - Configuración Optimizada 20x Más Grande e Inteligente
|
| 2 |
+
|
| 3 |
+
model:
|
| 4 |
+
vocab_size: 8000 # 2x más vocabulario
|
| 5 |
+
d_model: 1024 # 2x dimensión (512 → 1024)
|
| 6 |
+
n_layers: 24 # 3x capas (8 → 24)
|
| 7 |
+
n_heads: 16 # 2x cabezas (8 → 16)
|
| 8 |
+
d_ff: 4096 # 4x d_model
|
| 9 |
+
max_seq_len: 2048 # 4x contexto (512 → 2048)
|
| 10 |
+
dropout: 0.15 # Dropout optimizado
|
| 11 |
+
use_swiglu: true # Mejor activación
|
| 12 |
+
use_flash_attention: true # Atención optimizada
|
| 13 |
+
use_confidence_scoring: true # Anti-alucinación
|
| 14 |
+
min_confidence: 0.3
|
| 15 |
+
|
| 16 |
+
training:
|
| 17 |
+
batch_size: 2 # Pequeño para modelo grande
|
| 18 |
+
accumulation_steps: 16 # Effective batch = 32
|
| 19 |
+
epochs: 25 # 25 épocas como pediste
|
| 20 |
+
learning_rate: 0.0002 # LR bajo para estabilidad
|
| 21 |
+
min_lr: 0.000005
|
| 22 |
+
weight_decay: 0.15 # Regularización fuerte
|
| 23 |
+
max_grad_norm: 0.5
|
| 24 |
+
num_threads: 4
|
| 25 |
+
save_every: 5 # Guardar cada 5 épocas
|
| 26 |
+
|
| 27 |
+
# Early stopping (para no perder info)
|
| 28 |
+
patience: 10 # Muy paciente (espera 10 épocas sin mejora)
|
| 29 |
+
min_delta: 0.0003 # Mejora mínima aceptable
|
| 30 |
+
|
| 31 |
+
# Learning rate
|
| 32 |
+
warmup_steps: 500
|
| 33 |
+
use_lr_scheduler: true
|
| 34 |
+
|
| 35 |
+
# Regularización
|
| 36 |
+
label_smoothing: 0.15
|
| 37 |
+
use_eos_loss_weight: true
|
| 38 |
+
eos_weight: 3.0
|
| 39 |
+
|
| 40 |
+
# Optimizaciones GPU
|
| 41 |
+
use_gradient_checkpointing: true # Ahorra VRAM
|
| 42 |
+
use_fp16: true # Mixed precision
|
| 43 |
+
|
| 44 |
+
data:
|
| 45 |
+
corpus_path: corpus/mtp_mini_corpus.jsonl
|
| 46 |
+
min_text_length: 100
|
| 47 |
+
max_text_length: 4000
|
| 48 |
+
validation_split: 0.2 # 20% para validación
|
| 49 |
+
|
| 50 |
+
# Augmentación
|
| 51 |
+
use_augmentation: true
|
| 52 |
+
augmentation_prob: 0.4
|
| 53 |
+
|
| 54 |
+
generation:
|
| 55 |
+
default_max_tokens: 300
|
| 56 |
+
default_temperature: 0.65
|
| 57 |
+
default_top_k: 50
|
| 58 |
+
default_top_p: 0.9
|
| 59 |
+
default_repetition_penalty: 1.2
|
| 60 |
+
min_response_length: 30
|
| 61 |
+
|
| 62 |
+
# Anti-alucinación
|
| 63 |
+
use_perplexity_filter: true
|
| 64 |
+
max_perplexity: 80.0
|
| 65 |
+
use_entropy_threshold: true
|
| 66 |
+
max_entropy: 4.0
|
| 67 |
+
|
| 68 |
+
# Control de calidad
|
| 69 |
+
use_confidence_filter: true
|
| 70 |
+
min_confidence_threshold: 0.3
|
| 71 |
+
|
| 72 |
+
stop_sequences:
|
| 73 |
+
- "###"
|
| 74 |
+
- "\n\n\n\n"
|
| 75 |
+
- "Instrucción:"
|
| 76 |
+
- "Usuario:"
|
| 77 |
+
|
| 78 |
+
# Optimización de memoria
|
| 79 |
+
memory:
|
| 80 |
+
use_fp16: true
|
| 81 |
+
use_gradient_checkpointing: true
|
| 82 |
+
max_memory_gb: 14
|